# Nanbeige4.3-33B (slim)

> Nanbeige4.3-33B is the Nanbeige LLM Lab's open-weight dense model for agentic coding and office work, 33B parameters on the Llama architecture with a 512K-token context, thinking that switches per request and tool calls in an XML format.

- Full: https://www.anchorterminal.com/tools/nanbeige-4-3-33b.md (~9,150 tokens) · this version ~2,030 tokens · JSON https://www.anchorterminal.com/tools/nanbeige-4-3-33b.json · canonical https://www.anchorterminal.com/tools/nanbeige-4-3-33b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-25 · pricing https://www.anchorterminal.com/tools/nanbeige-4-3-33b#pricing · reviews https://www.anchorterminal.com/tools/nanbeige-4-3-33b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #56 of 111 open-weight models · #48 in Open-weight models · not agent-ready · confidence low**


Assessment: A dense 33B Llama-architecture cousin for long agentic coding and office tasks, with per-request thinking, XML tool calls that stock vLLM and SGLang parse, and 512K of context. It needs about 96 GB at BF16, as Nanbeige ships no quantised build. Its agentic results so far are Nanbeige's own, with no safety results, training data notes or licence file.

## Facts

- Kind: Open-weight model · vendor: Nanbeige LLM Lab (BOSS Zhipin) · category: Open-weight models · legal entity: BOSS Zhipin (Kanzhun Limited), unconfirmed · provenance 53/100
- Packages: huggingface `Nanbeige/Nanbeige4.3-33B`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0 (Hub metadata only; the repository has no LICENSE file)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Nanbeige/Nanbeige4.3-33B, read from the card, the files and their metadata, the chat template and the runtimes' own code on 25 October 2026, with the Base repository as a note. Nothing was downloaded or run
- Model: LlamaForCausalLM, 33,118,977,024 BF16 parameters by the Hub's count, 52 layers, hidden size 7,168, 56 attention heads with 8 KV heads (grouped-query attention), head size 128, vocabulary 166,912, untied embeddings. Standard Llama tensor names, no custom code and no auto_map. Text in and text out; the template answers images, audio and video with a note that the model can't read them
- Files: 66.26 GB in all, 14 BF16 safetensors shards (66.24 GB), tokenizer.json and a SentencePiece tokenizer.model, tokenizer_config.json with the chat template, configs and three figures. No pickle files, no Python files, no LICENSE file. SHA-256 on every LFS file
- Licence: Apache 2.0 in the card's metadata (license apache-2.0) and on the Base card. Neither repository has a LICENSE file and the card text has no licence section, so the terms rest on the metadata tag. No use policy or other terms found
- Runtimes: The card gives Transformers, vLLM and SGLang commands. Released vLLM 0.32.0, SGLang 0.5.22, Transformers 5.20.2 and llama.cpp b11888 all carry LlamaForCausalLM, and vLLM and SGLang carry the qwen3 reasoning parser and qwen3_coder tool parser the card names. No Ollama library entry, no LM Studio catalogue entry and no ggml-org or lmstudio-community build on 25 October 2026
- Context: 524,288 tokens, max_position_embeddings in config.json with rope_theta 120,000,000 and no rope scaling, and the card's 512K. The Base model is 131,072 tokens with rope_theta 10,000,000. The card doesn't say at what length the model was trained
- Tools and reasoning: Passing tools switches the template to a tool prompt. tool_call_format xml (the default, which Nanbeige recommends) asks for <tool_call><function=name><parameter=...> blocks, read back by the qwen3_coder parser in vLLM and SGLang; json asks for Hermes-style JSON in <tool_call> tags. The template renders several calls per turn and groups tool results into one user turn. enable_thinking switches thinking per request (on by default), the qwen3 reasoning parser splits it out, and preserve_thinking keeps earlier turns' reasoning, which Nanbeige recommends for agent and coding work. No thinking budget documented
- Hardware: Not stated. The card's vLLM and SGLang commands use tensor parallelism of 4. The BF16 files need about 80 GB with a fifth added, so 96 GB of memory, within the 128 GB reference machine's GPU share
- Training data: Not described on either card, with no cut-off. Nanbeige's earlier technical reports covered earlier models only

- Scores: Capability pending, Reach 55, Risk 60, Agent interface 85, Documentation and upkeep 56 · total over the 7 assessed categories
- Why:  · Reach, Memory earns 10, since the only build from Nanbeige or a runtime's library is 66.26 GB of BF16, about 80 GB with a fifth added, which fits t… · Risk, File safety earns 25, safetensors only and no repository code, so the card's trust_remote_code has nothing to run. · Agent interface, The chat template earns 20, Jinja in tokenizer_config.json. · Documentation and upkeep, The model card earns 11, with intended use, parameters, context and languages but no limitations, architecture or cut-off.
- Sources: 19, open questions: 10, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nanbeige-4-3-33b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nanbeige-4-3-33b.svg` or a link to https://www.anchorterminal.com/tools/nanbeige-4-3-33b from a page on zhipin.com or one of its subdomains, or a page under huggingface.co/nanbeige, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Serve with `vllm serve Nanbeige/Nanbeige4.3-33B --enable-auto-tool-choice --reasoning-parser qwen3 --tool-call-parser qwen3_coder`, as the card says, or SGLang with the same two parsers
2. Keep `tool_call_format` at its default `xml`, since qwen3_coder parses that format; the `json` option writes Hermes-style JSON calls, which need a different parser such as vLLM's hermes
3. Pass `chat_template_kwargs` `{"preserve_thinking": true}` for multi-turn tool use and coding, and `{"enable_thinking": false}` to skip reasoning, as the card recommends
4. Use temperature 1.0, top_p 0.95 and up to 65,536 new tokens for agent work, and temperature 0.6 for reasoning and chat, per the card
5. Send your own system prompt, since without one the template injects a Chinese identity line, or a Chinese tool-expert prompt when tools are passed
6. Send one system message at the start only, as the template raises an error for a later one
7. In plain Transformers pass `use_cache=True` to generate, since config.json and generation_config.json set it false; trust_remote_code isn't needed

## Connect

```bash
# vLLM 0.32.0 or SGLang 0.5.22, as the card's commands (the card uses four GPUs with tensor parallelism)
hf download Nanbeige/Nanbeige4.3-33B --local-dir Nanbeige4.3-33B
vllm serve ./Nanbeige4.3-33B --served-model-name nanbeige --tensor-parallel-size 4 \
  --enable-auto-tool-choice --reasoning-parser qwen3 --tool-call-parser qwen3_coder
# or
python -m sglang.launch_server --model-path ./Nanbeige4.3-33B --tp-size 4 \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder
```

```bash
curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model": "nanbeige", "messages": [{"role": "user", "content": "Which number is bigger, 9.11 or 9.8?"}],
       "temperature": 0.6, "top_p": 0.95, "chat_template_kwargs": {"enable_thinking": true}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)